Title stata.com fp Fractional polynomial regression Syntax Menu Description Options for fp Options for fp generate Remarks and examples Stored results Methods and formulas Acknowledgment References Also see Syntax Estimation fp <term> [, est options ] : est cmd est cmd may be almost any estimation command that stores the e(ll) result. To confirm whether fp works with a specific est cmd, see the documentation for that est cmd. Instances of <term> (with the angle brackets) that occur within est cmd are replaced in est cmd by a varlist containing the fractional powers of the variable term. These variables will be named term 1, term 2,.... fp performs est cmd with this substitution, fitting a fractional polynomial regression in term. est cmd in either this or the following syntax may not contain other prefix commands; see [U] 11.1.10 Prefix commands. Estimation (alternate syntax) fp <term>(varname) [, est options ] : est cmd Use this syntax to specify that fractional powers of varname are to be calculated. The fractional polynomial power variables will still be named term 1, term 2,.... Replay estimation results fp [, replay options ] Create specified fractional polynomial power variables fp generate [ type ] [ newvar = ] varname^(numlist) [ if ] [ in ] [, gen options ] 1
2 fp Fractional polynomial regression est options Description Main Search powers(# #... #) powers to be searched; default is powers(-2-1 -.5 0.5 1 2 3) dimension(#) maximum degree of fractional polynomial; default is dimension(2) Or specify fp(# #... #) And then specify any of these options use specified fractional polynomial Options classic replace all scale(# a # b) scale center(# c) center zero catzero Reporting replay options perform automatic scaling and centering and omit comparison table replace existing fractional polynomial power variables named term 1, term 2,... generate term 1, term 2,... in all observations; default is in observations if esample() use (term+a)/b; default is to use variable term as is specify a and b automatically report centered-on-c results; default is uncentered results specify c to be the mean of (scaled) term set term 1, term 2,... to zero if scaled term 0; default is to issue an error message same as zero and include term 0 = (term 0) among fractional polynomial power variables specify how results are displayed replay options Reporting nocompare reporting options Description do not display model-comparison test results any options allowed by est cmd for replaying estimation results gen options Main replace scale(# a # b) scale center(# c) center zero catzero Description replace existing fractional polynomial power variables named term 1, term 2,... use (term+a)/b; default is to use variable term as is specify a and b automatically report centered-on-c results; default is uncentered results specify c to be the mean of (scaled) term set term 1, term 2,... to zero if scaled term 0; default is to issue an error message same as zero and include term 0 = (term 0) among fractional polynomial power variables
Menu fp Fractional polynomial regression 3 fp Statistics > Linear models and related > Fractional polynomials > Fractional polynomial regression fp generate Statistics > Linear models and related > Fractional polynomials > Create fractional polynomial variables Description fp <term>: est cmd fits models with the best -fitting fractional polynomial substituted for <term> wherever it appears in est cmd. fp <weight>: regress mpg <weight> foreign would fit a regression model of mpg on a fractional polynomial in weight and (linear) foreign. By specifying option fp(), you may set the exact powers to be used. Otherwise, a search through all possible fractional polynomials up to the degree set by dimension() with the powers set by powers() is performed. fp without arguments redisplays the previous estimation results, just as typing est cmd would. You can type either one. fp will include a fractional polynomial comparison table. fp generate creates fractional polynomial power variables for a given set of powers. For instance, fp <weight>: regress mpg <weight> foreign might produce the fractional polynomial weight ( 2, 1) and store weight 2 in weight 1 and weight 1 in weight 2. Typing fp generate weight^(-2-1) would allow you to create the same variables in another dataset. See [R] mfp for multivariable fractional polynomial models. Options for fp Main powers(# #... #) specifies that a search be performed and details about the search provided. powers() works with the dimension() option; see below. The default is powers(-2-1 -.5 0.5 1 2 3). dimension(#) specifies the maximum degree of the fractional polynomial to be searched. The default is dimension(2). If the defaults for both powers() and dimension() are used, then the fractional polynomial could be any of the following 44 possibilities:
4 fp Fractional polynomial regression term ( 2) term ( 1).. term (3) term ( 2), term ( 2) term ( 2), term ( 1). term ( 2), term (3) term ( 1), term ( 2). term (3), term (3) fp(# #... #) specifies that no search be performed and that the fractional polynomial specified be used. fp() is an alternative to powers() and dimension(). Options classic performs automatic scaling and centering and omits the comparison table. Specifying classic is equivalent to specifying scale, center, and nocompare. replace replaces existing fractional polynomial power variables named term 1, term 2,.... all specifies that term 1, term 2,... be filled in for all observations in the dataset rather than just for those in e(sample). scale(# a # b) specifies that term be scaled in the way specified, namely, that (term+a)/b be calculated. All values of scaled term are required to be greater than zero unless you specify options zero or catzero. Values should not be too large or too close to zero, because by default, cubic powers and squared reciprocal powers will be considered. When scale(a b) is specified, values in the variable term are not modified; fp merely remembers to scale the values whenever powers are calculated. You will probably not use scale(a b) for values of a and b that you create yourself, although you could. It is usually easier just to generate a scaled variable. For instance, if term is age, and age in your data is required to be greater than or equal to 20, you might generate an age5 variable, for use as term:. generate age5 = (age-19)/5 scale(a b) is useful when you previously fit a model using automatic scaling (option scale) in one dataset and now want to create the fractional polynomials in another. In the first dataset, fp with scale added notes to the dataset concerning the values of a and b. You can see them by typing. notes You can then use fp generate, scale(a b) in the second dataset. The default is to use term as it is used in calculating fractional powers; thus term s values are required to be greater than zero unless you specify options zero or catzero. Values should not be too large, because by default, cubic powers will be considered. scale specifies that term be scaled to be greater than zero and not too large in calculating fractional powers. See Scaling for more details. When scale is specified, values in the variable term are not modified; fp merely remembers to scale the values whenever powers are calculated.
fp Fractional polynomial regression 5 center(# c) reports results for the fractional polynomial in (scaled) term, centered on c. The default is to perform no centering. term (p1,p2,...,pm) -c (p1,p2,...,pm) is reported. This makes the constant coefficient (intercept) easier to interpret. See Centering for more details. center performs center(c), where c is the mean of (scaled) term. zero and catzero specify how nonpositive values of term are to be handled. By default, nonpositive values of term are not allowed, because we will be calculating natural logarithms and fractional powers of term. Thus an error message is issued. zero sets the fractional polynomial value to zero for nonpositive values of (scaled) term. catzero sets the fractional polynomial value to zero for nonpositive values of (scaled) term and includes a dummy variable indicating where nonpositive values of (scaled) term appear in the model. Reporting nocompare suppresses display of the comparison tests. reporting options are any options allowed by est cmd for replaying estimation results. Options for fp generate Main replace replaces existing fractional polynomial power variables named term 1, term 2,.... scale(# a # b) specifies that term be scaled in the way specified, namely, that (term+a)/b be calculated. All values of scaled term are required to be greater than zero unless you specify options zero or catzero. Values should not be too large or too close to zero, because by default, cubic powers and squared reciprocal powers will be considered. When scale(a b) is specified, values in the variable term are not modified; fp merely remembers to scale the values whenever powers are calculated. You will probably not use scale(a b) for values of a and b that you create yourself, although you could. It is usually easier just to generate a scaled variable. For instance, if term is age, and age in your data is required to be greater than or equal to 20, you might generate an age5 variable, for use as term:. generate age5 = (age-19)/5 scale(a b) is useful when you previously fit a model using automatic scaling (option scale) in one dataset and now want to create the fractional polynomials in another. In the first dataset, fp with scale added notes to the dataset concerning the values of a and b. You can see them by typing. notes You can then use fp generate, scale(a b) in the second dataset. The default is to use term as it is used in calculating fractional powers; thus term s values are required to be greater than zero unless you specify options zero or catzero. Values should not be too large, because by default, cubic powers will be considered. scale specifies that term be scaled to be greater than zero and not too large in calculating fractional powers. See Scaling for more details. When scale is specified, values in the variable term are not modified; fp merely remembers to scale the values whenever powers are calculated.
6 fp Fractional polynomial regression center(# c) reports results for the fractional polynomial in (scaled) term, centered on c. The default is to perform no centering. term (p1,p2,...,pm) -c (p1,p2,...,pm) is reported. This makes the constant coefficient (intercept) easier to interpret. See Centering for more details. center performs center(c), where c is the mean of (scaled) term. zero and catzero specify how nonpositive values of term are to be handled. By default, nonpositive values of term are not allowed, because we will be calculating natural logarithms and fractional powers of term. Thus an error message is issued. zero sets the fractional polynomial value to zero for nonpositive values of (scaled) term. catzero sets the fractional polynomial value to zero for nonpositive values of (scaled) term and includes a dummy variable indicating where nonpositive values of (scaled) term appear in the model. Remarks and examples stata.com Remarks are presented under the following headings: Fractional polynomial regression Scaling Centering Examples Fractional polynomial regression Regression models based on fractional polynomial functions of a continuous covariate are described by Royston and Altman (1994). Fractional polynomials increase the flexibility afforded by the family of conventional polynomial models. Although polynomials are popular in data analysis, linear and quadratic functions are limited in their range of curve shapes, whereas cubic and higher-order curves often produce undesirable artifacts such as edge effects and waves. Fractional polynomials differ from regular polynomials in that 1) they allow logarithms, 2) they allow noninteger powers, and 3) they allow powers to be repeated. We will write a fractional polynomial in x as x (p1,p2,...,pm) β We will write x (p) to mean a regular power except that x (0) is to be interpreted as meaning ln(x) rather than x (0) = 1. Then if there are no repeated powers in (p 1, p 2,..., p m ), x (p1,p2,...,pm) β = β 0 + β 1 x (p1) + β 2 x (p2) + + β m x (pm) Powers are allowed to repeat in fractional polynomials. Each time a power repeats, it is multiplied by another ln(x). As an extreme case, consider the fractional polynomial with all-repeated powers, say, m of them, x (p,p,...,p) β = β 0 + β 1 x (p) + β 2 x (p) ln(x) + + β m x (p) {ln(x)} m 1
fp Fractional polynomial regression 7 Thus the fractional polynomial x (0,0,2) β would be x (0,0,2) β = β 0 + β 1 x (0) + β 2 x (0) ln(x) + β 3 x (2) = β 0 + β 1 ln(x) + β 2 {ln(x)} 2 + β 3 x 2 With this definition, we can obtain a much wider range of shapes than can be obtained with regular polynomials. The following graphs appeared in Royston and Sauerbrei (2008, sec. 4.5). The first graph shows the shapes of differing fractional polynomials. 2 1 0.5 0 0.5 1 2 3
8 fp Fractional polynomial regression The second graph shows some of the curve shapes available with different βs for the degree-2 fractional polynomial, x ( 2,2). In modeling a fractional polynomial, Royston and Sauerbrei (2008) recommend choosing powers from among { 2, 1, 0.5, 0, 0.5, 1, 2, 3}. By default, fp chooses powers from this set, but other powers can be explicitly specified in the powers() option. fp <term>: est cmd fits models with the terms of the best-fitting fractional polynomial substituted for <term> wherever it appears in est cmd. We will demonstrate with auto.dta, which contains repair records and other information about a variety of vehicles in 1978. We use fp to find the best fractional polynomial in automobile weight (lbs.) (weight) for the linear regression of miles per gallon (mpg) on weight and an indicator of whether the vehicle is foreign (foreign). By default, fp will fit degree-2 fractional polynomial (FP2) models and choose the fractional powers from the set { 2, 1, 0.5, 0, 0.5, 1, 2, 3}. Because car weight is measured in pounds and will have a cubic transformation applied to it, we shrink it to a smaller scale before estimation by dividing by 1,000. We modify the existing weight variable for conciseness and to facilitate the comparison of tables. When applying a data transformation in practice, rather than modifying the existing variables, you should create new variables that hold the transformed values.
fp Fractional polynomial regression 9. use http://www.stata-press.com/data/r13/auto (1978 Automobile Data). replace weight = weight/1000 weight was int now float (74 real changes made). fp <weight>: regress mpg <weight> foreign (fitting 44 models) (...10%...20%...30%...40%...50%...60%...70%...80%...90%...100%) Fractional polynomial comparisons: weight df Deviance Res. s.d. Dev. dif. P(*) Powers omitted 0 456.347 5.356 75.216 0.000 linear 1 388.366 3.407 7.236 0.082 1 m = 1 2 381.806 3.259 0.675 0.733 -.5 m = 2 4 381.131 3.268 0.000 -- -2-2 (*) P = sig. level of model with m = 2 based on F with 68 denominator dof. Source SS df MS Number of obs = 74 F( 3, 70) = 52.95 Model 1696.05949 3 565.353163 Prob > F = 0.0000 Residual 747.399969 70 10.6771424 R-squared = 0.6941 Adj R-squared = 0.6810 Total 2443.45946 73 33.4720474 Root MSE = 3.2676 mpg Coef. Std. Err. t P> t [95% Conf. Interval] weight_1 15.88527 20.60329 0.77 0.443-25.20669 56.97724 weight_2 127.9349 47.53106 2.69 0.009 33.13723 222.7326 foreign -2.222515 1.053782-2.11 0.039-4.324218 -.1208131 _cons 3.705981 3.367949 1.10 0.275-3.011182 10.42314 fp begins by showing the model-comparison table. This table shows the best models of each examined degree, obtained by searching through all possible power combinations. The fractional powers of the models are shown in the Powers column. A separate row is provided for the linear fractional polynomial because it is often the default used when including a predictor in the model. The null model does not include any fractional polynomial terms for weight. The df column shows the count of the additional parameters used in each model beyond the quantity of parameters used in the null model. The model deviance, which we define as twice the negative log likelihood, is given in the Deviance column. The difference of the model deviance from the deviance of the model with the lowest deviance is given in the Dev. dif. column. The p-value for the partial F test comparing the models and the lowest deviance model is given in the P(*) column. An estimate of the residual standard error is given in the Res. Sd. column. Under linear regression, a partial F test is used in the model-comparison table. In other settings, a likelihood-ratio test is performed. Then a χ 2 statistic is reported. Under robust variance estimation and some other cases (see [R] lrtest), the likelihood-ratio test cannot be performed. When the likelihood-ratio test cannot be performed on the model specified in est cmd, fp still reports the model-comparison table, but the comparison tests are not performed. fp reports the best model as the model with the lowest deviance; however, users may choose a more efficient model based on the comparison table. They may choose the lowest degree model that the partial F test (or likelihood-ratio test) fails to reject in favor of the lowest deviance model. After the comparison table, the results of the estimation command for the lowest deviance model are shown. Here the best model has terms weight ( 2, 2). However, based on the model-comparison
10 fp Fractional polynomial regression table, we can reject the model without weight and the linear model at the 0.1 significance level. We fail to reject the m = 1 model at any reasonable level. We will choose the FP1 model, which includes weight (.5). We use fp again to estimate the parameters for this model. We use the fp() option to specify what powers we want to use; this option specifies that we do not want to perform a search for the best powers. We also specify the replace option to overwrite the previously created fractional polynomial power variables.. fp <weight>, fp(-.5) replace: regress mpg <weight> foreign -> regress mpg weight_1 foreign Source SS df MS Number of obs = 74 F( 2, 71) = 79.51 Model 1689.20865 2 844.604325 Prob > F = 0.0000 Residual 754.25081 71 10.6232508 R-squared = 0.6913 Adj R-squared = 0.6826 Total 2443.45946 73 33.4720474 Root MSE = 3.2593 mpg Coef. Std. Err. t P> t [95% Conf. Interval] weight_1 66.89665 6.021749 11.11 0.000 54.88963 78.90368 foreign -2.095622 1.043513-2.01 0.048-4.176329 -.0149157 _cons -17.58651 3.397992-5.18 0.000-24.36192-10.81111 Alternatively, we can use fp generate to create the fractional polynomial variable corresponding to weight (.5) and then use regress. We store weight (.5) in the new variable wgt nsqrt.. fp generate wgt_nsqrt=weight^(-.5). regress mpg wgt_nsqrt foreign Source SS df MS Number of obs = 74 F( 2, 71) = 79.51 Model 1689.20874 2 844.604371 Prob > F = 0.0000 Residual 754.250718 71 10.6232495 R-squared = 0.6913 Adj R-squared = 0.6826 Total 2443.45946 73 33.4720474 Root MSE = 3.2593 mpg Coef. Std. Err. t P> t [95% Conf. Interval] wgt_nsqrt_1 66.89665 6.021748 11.11 0.000 54.88963 78.90368 foreign -2.095622 1.043513-2.01 0.048-4.176328 -.0149155 _cons -17.58651 3.397991-5.18 0.000-24.36191-10.81111 Scaling Fractional polynomials are only defined for positive term variables. By default, fp will assume that the variable x is positive and attempt to compute fractional powers of x. If the positive value assumption is incorrect, an error will be reported and estimation will not be performed. If the values of the variable are too large or too small, the reported results of fp may be difficult to interpret. By default, cubic powers and squared reciprocal powers will be considered in the search for the best fractional polynomial in term. We can scale the variable x to 1) make it positive and 2) ensure its magnitude is not too large or too small. Suppose you have data on hospital patients with age as a fractional polynomial variable of interest. age is required to be greater than or equal to 20, so you might generate an age5 variable by typing
fp Fractional polynomial regression 11. generate age5 = (age-19)/5 A unit change in age5 is equivalent to a five-year change in age, and the minimum value of age5 is 1/5 instead of 20. In the automobile example of Fractional polynomial regression, our term variable was automobile weight (lbs.). Cars weigh in the thousands of pounds, so cubing their weight figures results in large numbers. We prevented this from being a problem by shrinking the weight by 1,000; that is, we typed. replace weight = weight/1000 Calendar year is another type of variable that can have a problematically large magnitude. We can shrink this by dividing by 10, making a unit change correspond to a decade.. generate decade = calendar_year/10 You may also have a variable that measures deviation from zero. Perhaps x has already been demeaned and is symmetrically about zero. The fractional polynomial in x will be undefined for half of its domain. We can shift the location of x, making it positive by subtracting its minimum and adding a small number to it. Suppose x ranges from 4 to 4; we could use. generate newx = x+5 Rescaling ourselves provides easily communicated results. We can tell exactly how the scaling was performed and how it should be performed in similar applications. Alternatively, fp can scale the fractional polynomial variable so that its values are positive and the magnitude of the values are not too large. This can be done automatically or by directly specifying the scaling values. Scaling can be automatically performed with fp by specifying the scale option. If term has nonpositive values, the minimum value of term is subtracted from each observation of term. In this case, the counting interval, the minimum distance between the sorted values of term, is also added to each observation of term. After adjusting the location of term so that its minimum value is positive, creating term, automatic scaling will divide each observation of term by a power of ten. The exponent of this scaling factor is given by p = log 10 {max(term ) min(term )} p = sign(p)floor ( p ) Rather than letting fp automatically choose the scaling of term, you may specify adjustment and scale factors a and b by using the scale(a b) option. Fractional powers are then calculated using the (term+a)/b values. When scale or scale(a b) is specified, values in the variable term are not modified; fp merely remembers to scale the values whenever powers are calculated. In addition to fp, both scale and scale(a b) may be used with fp generate. You will probably not use scale(a b) with fp for values of a and b that you create yourself, although you could. As we demonstrated earlier, it is usually easier just to generate a scaled variable. scale(a b) is useful when you previously fit a model using scale in one dataset and now want to create the fractional polynomials in another. In the first dataset, fp with scale added notes to the dataset concerning the values of a and b. You can see them by typing. notes
12 fp Fractional polynomial regression You can then use fp generate, scale(a b) in the second dataset. When you apply the scaling rules of a previously fit model to new data with the scale(a b) option, it is possible that the scaled term may have nonpositive values. fp will be unable to calculate the fractional powers of the term in this case and will issue an error. The options zero and catzero cause fp and fp generate to output zero values for each fractional polynomial variable when the input (scaled) fractional polynomial variable is nonpositive. Specifying catzero causes a dummy variable indicating nonpositive values of the (scaled) fractional polynomial variable to be included in the model. A detailed example of the use of catzero and zero is shown in example 3 below. Using the scaling options, we can fit our previous model again using the auto.dta. We specify scale(0 1000) so that fp will shrink the magnitude of weight in estimating the regression. This is done for demonstration purposes because our scaling rule is simple. As mentioned before, in practice, you would probably only use scale(a b) when applying the scaling rules from a previous analysis. Allowing fp to scale does have the advantage of not altering the original variable, weight.. use http://www.stata-press.com/data/r13/auto, clear (1978 Automobile Data). fp <weight>, fp(-.5) scale(0 1000): regress mpg <weight> foreign -> regress mpg weight_1 foreign Source SS df MS Number of obs = 74 F( 2, 71) = 79.51 Model 1689.20861 2 844.604307 Prob > F = 0.0000 Residual 754.250846 71 10.6232514 R-squared = 0.6913 Adj R-squared = 0.6826 Total 2443.45946 73 33.4720474 Root MSE = 3.2593 mpg Coef. Std. Err. t P> t [95% Conf. Interval] weight_1 66.89665 6.021749 11.11 0.000 54.88963 78.90368 foreign -2.095622 1.043513-2.01 0.048-4.176329 -.0149159 _cons -17.58651 3.397992-5.18 0.000-24.36192-10.81111 The scaling is clearly indicated in the variable notes for the generated variable weight 1.. notes weight_1 weight_1: 1. fp term 1 of x^(-.5), where x is weight scaled. 2. Scaling was user specified: x = (weight+a)/b where a=0 and b=1000 3. Fractional polynomial variables created by fp <weight>, fp(-.5) scale(0 1000): regress mpg <weight> foreign 4. To re-create the fractional polynomial variables, for instance, in another dataset, type fp gen double weight^(-.5), scale(0 1000) Centering The fractional polynomial of term, centered on c is ( term (p1,...,pm) c (p1,...,pm)) β The intercept of a centered fractional polynomial can be interpreted as the effect at zero for all the covariates. When we center the fractional polynomial terms using c, the intercept is now interpreted as the effect at term = c and zero values for the other covariates.
fp Fractional polynomial regression 13 Suppose we wanted to center the fractional polynomial of x with powers (0, 0, 2) at x = c. ( x (0,0,2) c (0,0,2)) β ( = β 0 + β 1 x (0) c (0)) } + β 2 {x (0) ln(x) c (0) ln(c) + β 3 (x (2) c (2)) = β 0 + β 1 {ln(x) ln(c)} + β 2 [ {ln(x)} 2 {ln(c)} 2] + β 3 ( x 2 c 2) When center is specified, fp centers based on the sample mean of (scaled) term. A previously chosen value for centering, c, may also be specified in center(c). This would be done when applying the results of a previous model fitting to a new dataset. The center and center(c) options may be used in fp or fp generate. Returning to the model of mileage per gallon based on automobile weight and foreign origin, we refit the model with the fractional polynomial of weight centered at its scaled mean.. use http://www.stata-press.com/data/r13/auto, clear (1978 Automobile Data). fp <weight>, fp(-.5) scale(0 1000) center: regress mpg <weight> foreign -> regress mpg weight_1 foreign Source SS df MS Number of obs = 74 F( 2, 71) = 79.51 Model 1689.20861 2 844.604307 Prob > F = 0.0000 Residual 754.250846 71 10.6232514 R-squared = 0.6913 Adj R-squared = 0.6826 Total 2443.45946 73 33.4720474 Root MSE = 3.2593 mpg Coef. Std. Err. t P> t [95% Conf. Interval] weight_1 66.89665 6.021749 11.11 0.000 54.88963 78.90368 foreign -2.095622 1.043513-2.01 0.048-4.176329 -.0149159 _cons 20.91163.4624143 45.22 0.000 19.9896 21.83366 Note that the coefficients for weight 1 and foreign do not change. Only the intercept cons changes. It can be interpreted as the estimated average miles per gallon of an American-made car of average weight.
14 fp Fractional polynomial regression Examples Example 1: Linear regression Consider the serum immunoglobulin G (IgG) dataset from Isaacs et al. (1983), which consists of 298 independent observations in young children. The dependent variable sqrtigg is the square root of the IgG concentration, and the independent variable age is the age of each child. (Preliminary Box Cox analysis shows that a square root transformation removes the skewness in IgG.) The aim is to find a model that accurately predicts the mean of sqrtigg given age. We use fp to find the best FP2 model (the default option). We specify center for automatic centering. The age of each child is small in magnitude and positive, so we do not use the scaling options of fp or scale ourselves.. use http://www.stata-press.com/data/r13/igg, clear (Immunoglobulin in children). fp <age>, scale center: regress sqrtigg <age> (fitting 44 models) (...10%...20%...30%...40%...50%...60%...70%...80%...90%...100%) Fractional polynomial comparisons: age df Deviance Res. s.d. Dev. dif. P(*) Powers omitted 0 427.539 0.497 108.090 0.000 linear 1 337.561 0.428 18.113 0.000 1 m = 1 2 327.436 0.421 7.987 0.020 0 m = 2 4 319.448 0.416 0.000 -- -2 2 (*) P = sig. level of model with m = 2 based on F with 293 denominator dof. Source SS df MS Number of obs = 298 F( 2, 295) = 64.49 Model 22.2846976 2 11.1423488 Prob > F = 0.0000 Residual 50.9676492 295.172771692 R-squared = 0.3042 Adj R-squared = 0.2995 Total 73.2523469 297.246640898 Root MSE =.41566 sqrtigg Coef. Std. Err. t P> t [95% Conf. Interval] age_1 -.1562156.027416-5.70 0.000 -.2101713 -.10226 age_2.0148405.0027767 5.34 0.000.0093757.0203052 _cons 2.283145.0305739 74.68 0.000 2.222974 2.343315 The new variables created by fp contain the best-fitting fractional polynomial powers of age, as centered by fp. For example, age 1 is centered by subtracting the mean of age raised to the power 2. The variables created by fp and fp generate are centered or scaled as specified by the user, which is reflected in the estimated regression coefficients and intercept. Centering does have its advantages (see Centering earlier in this entry). By default, fp will not perform scaling or centering. For a more detailed discussion, see Royston and Sauerbrei (2008, sec. 4.11). The fitted curve has an asymmetric S shape. The best model has powers ( 2, 2) and deviance 319.448. We reject lesser degree models: the null, linear, and natural log power models at the 0.05 level. As many as 44 models have been fit in the search for the best powers. Now let s look at models of degree 4. The highest allowed degree is specified in dimension(). We overwrite the previously generated fractional polynomial power variables by including replace.
fp Fractional polynomial regression 15. fp <age>, dimension(4) center replace: regress sqrtigg <age> (fitting 494 models) (...10%...20%...30%...40%...50%...60%...70%...80%...90%...100%) Fractional polynomial comparisons: age df Deviance Res. s.d. Dev. dif. P(*) Powers omitted 0 427.539 0.497 109.795 0.000 linear 1 337.561 0.428 19.818 0.007 1 m = 1 2 327.436 0.421 9.692 0.149 0 m = 2 4 319.448 0.416 1.705 0.798-2 2 m = 3 6 319.275 0.416 1.532 0.476-2 1 1 m = 4 8 317.744 0.416 0.000 -- 0 3 3 3 (*) P = sig. level of model with m = 4 based on F with 289 denominator dof. Source SS df MS Number of obs = 298 F( 4, 293) = 32.63 Model 22.5754541 4 5.64386353 Prob > F = 0.0000 Residual 50.6768927 293.172958678 R-squared = 0.3082 Adj R-squared = 0.2987 Total 73.2523469 297.246640898 Root MSE =.41588 sqrtigg Coef. Std. Err. t P> t [95% Conf. Interval] age_1.8761824.1898721 4.61 0.000.5024962 1.249869 age_2 -.1922029.0684934-2.81 0.005 -.3270044 -.0574015 age_3.2043794.074947 2.73 0.007.0568767.3518821 age_4 -.0560067.0212969-2.63 0.009 -.097921 -.0140924 _cons 2.238735.0482705 46.38 0.000 2.143734 2.333736 It appears that the FP4 model is not significantly different from the other fractional polynomial models (at the 0.05 level). Let s compare the curve shape from the m = 2 model with that from a conventional quartic polynomial whose fit turns out to be significantly better than a cubic (not shown). We use the ability of fp both to generate the required powers of age, namely, (1, 2, 3, 4) for the quartic and ( 2, 2) for the second-degree fractional polynomial, and to fit the model. The fp() option is used to specify the powers. We use predict to obtain the fitted values of each regression. We fit both models with fp and graph the resulting curves with twoway scatter.
16 fp Fractional polynomial regression. fp <age>, center fp(1 2 3 4) replace: regress sqrtigg <age> -> regress sqrtigg age_1 age_2 age_3 age_4 Source SS df MS Number of obs = 298 F( 4, 293) = 32.65 Model 22.5835458 4 5.64588646 Prob > F = 0.0000 Residual 50.668801 293.172931061 R-squared = 0.3083 Adj R-squared = 0.2989 Total 73.2523469 297.246640898 Root MSE =.41585 sqrtigg Coef. Std. Err. t P> t [95% Conf. Interval] age_1 2.047831.4595962 4.46 0.000 1.143302 2.952359 age_2-1.058902.2822803-3.75 0.000-1.614456 -.5033479 age_3.2284917.0667591 3.42 0.001.0971037.3598798 age_4 -.0168534.0053321-3.16 0.002 -.0273475 -.0063594 _cons 2.240012.0480157 46.65 0.000 2.145512 2.334511. predict fit1 (option xb assumed; fitted values). label variable fit1 "Quartic". fp <age>, center fp(-2 2) replace: regress sqrtigg <age> -> regress sqrtigg age_1 age_2 Source SS df MS Number of obs = 298 F( 2, 295) = 64.49 Model 22.2846976 2 11.1423488 Prob > F = 0.0000 Residual 50.9676492 295.172771692 R-squared = 0.3042 Adj R-squared = 0.2995 Total 73.2523469 297.246640898 Root MSE =.41566 sqrtigg Coef. Std. Err. t P> t [95% Conf. Interval] age_1 -.1562156.027416-5.70 0.000 -.2101713 -.10226 age_2.0148405.0027767 5.34 0.000.0093757.0203052 _cons 2.283145.0305739 74.68 0.000 2.222974 2.343315. predict fit2 (option xb assumed; fitted values). label variable fit2 "FP 2". scatter sqrtigg fit1 fit2 age, c(. l l) m(o i i) msize(small) > lpattern(. -_.) ytitle("square root of IgG") xtitle("age, years")
fp Fractional polynomial regression 17 Square root of IgG 1 2 3 4 0 2 4 6 Age, years Square root of IgG FP 2 Quartic The quartic curve has an unsatisfactory wavy appearance that is implausible for the known behavior of IgG, the serum level of which increases throughout early life. The fractional polynomial curve (FP2) increases monotonically and is therefore biologically the more plausible curve. The two models have approximately the same deviance. Example 2: Cox regression Data from Smith et al. (1992) contain times to complete healing of leg ulcers in a randomized, controlled clinical trial of two treatments in 192 elderly patients. Several covariates were available, of which an important one is mthson, the number of months since the recorded onset of the ulcer. This time is recorded in whole months, not fractions of a month; therefore, some zero values are recorded. Because the response variable is time to an event of interest and some (in fact, about one-half) of the times are censored, using Cox regression to analyze the data is appropriate. We consider fractional polynomials in mthson, adjusting for four other covariates: age; ulcarea, the area of tissue initially affected by the ulcer; deepppg, a binary variable indicating the presence or absence of deep vein involvement; and treat, a binary variable indicating treatment type. We fit fractional polynomials of degrees 1 and 2 with fp. We specify scale to perform automatic scaling on mthson. This makes it positive and ensures that its magnitude is not too large. (See Scaling for more details.) The display option nohr is specified before the colon so that the coefficients and not the hazard ratios are displayed. The center option is specified to obtain automatic centering. age and ulcarea are also demeaned by using summarize and then subtracting the returned result r(mean). In Cox regression, there is no constant term, so we cannot see the effects of centering in the table of regression estimates. The effects would be present if we were to graph the baseline hazard or survival function because these functions are defined with all predictors set equal to 0. In these graphs, we will see the estimated baseline hazard or survival function under no deep vein involvement or treatment and under mean age, ulcer area, and number of months since the recorded onset of the ulcer.
18 fp Fractional polynomial regression. use http://www.stata-press.com/data/r13/legulcer1, clear (Leg ulcer clinical trial). stset ttevent, fail(cens) failure event: censored!= 0 & censored <. obs. time interval: (0, ttevent] exit on or before: failure 192 total observations 0 exclusions 192 observations remaining, representing 92 failures in single-record/single-failure data 13825 total analysis time at risk and under observation at risk from t = 0 earliest observed entry t = 0 last observed exit t = 206. qui sum age. replace age = age - r(mean) age was byte now float (192 real changes made). qui sum ulcarea. replace ulcarea = ulcarea - r(mean) ulcarea was int now float (192 real changes made). fp <mthson>, center scale nohr: stcox <mthson> age ulcarea deepppg treat (fitting 44 models) (...10%...20%...30%...40%...50%...60%...70%...80%...90%...100%) Fractional polynomial comparisons: mthson df Deviance Dev. dif. P(*) Powers omitted 0 754.345 17.636 0.001 linear 1 751.680 14.971 0.002 1 m = 1 2 738.969 2.260 0.323 -.5 m = 2 4 736.709 0.000 --.5.5 (*) P = sig. level of model with m = 2 based on chi^2 of dev. dif. Cox regression -- Breslow method for ties No. of subjects = 192 Number of obs = 192 No. of failures = 92 Time at risk = 13825 LR chi2(6) = 108.59 Log likelihood = -368.35446 Prob > chi2 = 0.0000 _t Coef. Std. Err. z P> z [95% Conf. Interval] mthson_1-2.81425.6996385-4.02 0.000-4.185516-1.442984 mthson_2 1.541451.4703143 3.28 0.001.6196521 2.46325 age -.0261111.0087983-2.97 0.003 -.0433556 -.0088667 ulcarea -.0017491.000359-4.87 0.000 -.0024527 -.0010455 deepppg -.5850499.2163173-2.70 0.007-1.009024 -.1610758 treat -.1624663.2171048-0.75 0.454 -.5879838.2630513 The best-fitting fractional polynomial of degree 2 has powers (0.5, 0.5) and deviance 736.709. However, this model does not fit significantly better than the fractional polynomial of degree 1 (at the 0.05 level), which has power 0.5 and deviance 738.969. We prefer the model with m = 1.
fp Fractional polynomial regression 19. fp <mthson>, replace center scale nohr fp(-.5): stcox <mthson> age ulcarea > deepppg treat -> stcox mthson_1 age ulcarea deepppg treat Cox regression -- Breslow method for ties No. of subjects = 192 Number of obs = 192 No. of failures = 92 Time at risk = 13825 LR chi2(5) = 106.33 Log likelihood = -369.48426 Prob > chi2 = 0.0000 _t Coef. Std. Err. z P> z [95% Conf. Interval] mthson_1.1985592.0493922 4.02 0.000.1017523.2953662 age -.02691.0087875-3.06 0.002 -.0441331 -.0096868 ulcarea -.0017416.0003482-5.00 0.000 -.0024241 -.0010591 deepppg -.5740759.2185134-2.63 0.009-1.002354 -.1457975 treat -.1798575.2175726-0.83 0.408 -.6062921.246577 The hazard for healing is much higher for patients whose ulcer is of recent onset than for those who have had an ulcer for many months. A more appropriate analysis of this dataset, if one wanted to model all the predictors, possibly with fractional polynomial functions, would be to use mfp; see [R] mfp. Example 3: Logistic regression The zero option permits fitting a fractional polynomial model to the positive values of a covariate, taking nonpositive values as zero. An application is the assessment of the effect of cigarette smoking as a risk factor. Whitehall 1 is an epidemiological study, which was examined in Royston and Sauerbrei (2008), of 18,403 male British Civil Servants employed in London. We examine the data collected in Whitehall 1 and use logistic regression to model the odds of death based on a fractional polynomial in the number of cigarettes smoked. Nonsmokers may be qualitatively different from smokers, so the effect of smoking (regarded as a continuous variable) may not be continuous between zero cigarettes and one cigarette. To allow for this possibility, we model the risk as a constant for the nonsmokers and as a fractional polynomial function of the number of cigarettes for the smokers, adjusted for age. The dependent variable all10 is an indicator of whether the individual passed away in the 10 years under study. cigs is the number of cigarettes consumed per day. After loading the data, we demean age and create a dummy variable, nonsmoker. We then use fp to fit the model.
20 fp Fractional polynomial regression. use http://www.stata-press.com/data/r13/smoking, clear (Smoking and mortality data). qui sum age. replace age = age - r(mean) age was byte now float (17260 real changes made). generate byte nonsmoker = cond(cigs==0, 1, 0) if cigs <.. fp <cigs>, zero: logit all10 <cigs> nonsmoker age (fitting 44 models) (...10%...20%...30%...40%...50%...60%...70%...80%...90%...100%) Fractional polynomial comparisons: cigs df Deviance Dev. dif. P(*) Powers omitted 0 9990.804 46.096 0.000 linear 1 9958.801 14.093 0.003 1 m = 1 2 9946.603 1.895 0.388 0 m = 2 4 9944.708 0.000 -- -1-1 (*) P = sig. level of model with m = 2 based on chi^2 of dev. dif. Logistic regression Number of obs = 17260 LR chi2(4) = 1029.03 Prob > chi2 = 0.0000 Log likelihood = -4972.3539 Pseudo R2 = 0.0938 all10 Coef. Std. Err. z P> z [95% Conf. Interval] cigs_1-1.285867.3358483-3.83 0.000-1.944117 -.6276162 cigs_2-1.982424.572109-3.47 0.001-3.103736 -.8611106 nonsmoker -1.223749.1119583-10.93 0.000-1.443183-1.004315 age.1194541.0045818 26.07 0.000.1104739.1284343 _cons -1.591489.1052078-15.13 0.000-1.797693-1.385286 Omission of the zero option would cause fp to halt with an error message because nonpositive covariate values (for example, values of cigs) are invalid unless the scale option is specified. A closely related approach involves the catzero option. Here we no longer need to have nonsmoker in the model, because fp creates its own dummy variable cigs 0 to indicate whether the individual does not smoke on that day.
fp Fractional polynomial regression 21. fp <cigs>, catzero replace: logit all10 <cigs> age (fitting 44 models) (...10%...20%...30%...40%...50%...60%...70%...80%...90%...100%) Fractional polynomial comparisons: cigs df Deviance Dev. dif. P(*) Powers omitted 0 10175.75 231.047 0.000 linear 2 9958.80 14.093 0.003 1 m = 1 3 9946.60 1.895 0.388 0 m = 2 5 9944.71 0.000 -- -1-1 (*) P = sig. level of model with m = 2 based on chi^2 of dev. dif. Logistic regression Number of obs = 17260 LR chi2(4) = 1029.03 Prob > chi2 = 0.0000 Log likelihood = -4972.3539 Pseudo R2 = 0.0938 all10 Coef. Std. Err. z P> z [95% Conf. Interval] cigs_0-1.223749.1119583-10.93 0.000-1.443183-1.004315 cigs_1-1.285867.3358483-3.83 0.000-1.944117 -.6276162 cigs_2-1.982424.572109-3.47 0.001-3.103736 -.8611106 age.1194541.0045818 26.07 0.000.1104739.1284343 _cons -1.591489.1052078-15.13 0.000-1.797693-1.385286 Under both approaches, the comparison table suggests that we can accept the FP1 model instead of the FP2 model. We estimate the parameters of the accepted model that is, the one that uses the natural logarithm of cigs with fp.. fp <cigs>, catzero replace fp(0): logit all10 <cigs> age -> logit all10 cigs_0 cigs_1 age Logistic regression Number of obs = 17260 LR chi2(3) = 1027.13 Prob > chi2 = 0.0000 Log likelihood = -4973.3016 Pseudo R2 = 0.0936 all10 Coef. Std. Err. z P> z [95% Conf. Interval] cigs_0.1883732.1553093 1.21 0.225 -.1160274.4927738 cigs_1.3469842.0543552 6.38 0.000.2404499.4535185 age.1194976.0045818 26.08 0.000.1105174.1284778 _cons -3.003767.1514909-19.83 0.000-3.300683-2.70685 The high p-value for cigs 0 in the output indicates that we cannot reject that there is no extra effect at zero for nonsmokers.
22 fp Fractional polynomial regression Stored results In addition to the results that est cmd stores, fp stores the following in e(): Scalars e(fp dimension) degree of fractional polynomial e(fp center mean) value used for centering or. e(fp scale a) value used for scaling or. e(fp scale b) value used for scaling or. e(fp compare df2) denominator degree of freedom in F test Macros e(fp cmd) e(fp cmdline) e(fp variable) e(fp terms) e(fp gen cmdline) e(fp catzero) e(fp zero) e(fp compare type) Matrices e(fp fp) e(fp compare) e(fp compare stat) e(fp compare df1) e(fp compare fp) e(fp compare length) e(fp powers) fp generate stores the following in r(): fp, search(): or fp, powers(): full fp command as typed fractional polynomial variable generated fp variables fp generate command to re-create e(fp terms) variables catzero, if specified zero, if specified F or chi2 powers used in fractional polynomial results of model comparisons F test statistics numerator degree of F test powers of comparison models encoded string for display of row titles powers that are searched Scalars r(fp center mean) value used for centering or. r(fp scale a) value used for scaling or. r(fp scale b) value used for scaling or. Macros r(fp cmdline) r(fp variable) r(fp terms) r(fp catzero) r(fp zero) Matrices r(fp fp) full fp generate command as typed fractional polynomial variable generated fp variables catzero, if specified zero, if specified powers used in fractional polynomial Methods and formulas The general definition of a fractional polynomial, accommodating possible repeated powers, may be written for functions H 1 (x),..., H m (x) of x > 0 as β 0 + where H 1 (x) = x (p1) and for j = 2,..., m, m β j H j (x) j=1 H j (x) = { x (p j) if p j p j 1 H j 1 (x) ln(x) if p j = p j 1
fp Fractional polynomial regression 23 For example, a fractional polynomial of degree 3 with powers (1, 3, 3) has H 1 (x) = x, H 2 (x) = x 3, and H 3 (x) = x 3 ln(x) and equals β 0 + β 1 x + β 2 x 3 + β 3 x 3 ln(x). We can express a fractional polynomial in vector notation by using H(x) = [H 1 (x),..., H d (x)]. We define x (p1,p2,...,pm) = [H(x), 1]. Under this notation, we can write x (1,3,3) β = β 0 + β 1 x + β 2 x 3 + β 3 x 3 ln(x) The fractional polynomial may be centered so that the intercept can be more easily interpreted. When centering the fractional polynomial of x at c, we subtract c (p1,p2,...,pm) from x (p1,p2,...,pm), where c (p1,p2,...,p d) = [H(x), 0]. The centered fractional polynomial is ( x (p1,...,p d) c (p1,...,p d) ) β The definition may be extended to allow x 0 values. For these values, the fractional polynomial is equal to the intercept β 0 or equal to a zero-offset term α 0 plus the intercept β 0. A fractional polynomial model of degree m is taken to have 2m + 1 degrees of freedom (df ): one for β 0 and one for each β j and its associated power. Because the powers in a fractional polynomial are chosen from a finite set rather than from the entire real line, the df defined in this way are approximate. The deviance D of a model is defined as 2 times its maximized log likelihood. For normal-errors models, we use the formula ( ) D = n 1 l + ln 2πRSS n where n is the sample size, l is the mean of the lognormalized weights (l = 0 if the weights are all equal), and RSS is the residual sum of squares as fit by regress. fp reports a table comparing fractional polynomial models of degree k < m with the degree m fractional polynomial model, which will have the lowest deviance. The p-values reported by fp are calculated differently for normal and nonnormal regressions. Let D k and D m be the deviances of the models with degrees k and m, respectively. For normal-errors models, a variance ratio F is calculated as F = n { 2 exp n 1 ( Dk D m n ) } 1 where n 1 is the numerator df, the quantity of the additional parameters that the degree m model has over the degree k model. n 2 is the denominator df and equals the residual degrees of freedom of the degree m model, minus the number of powers estimated, m. The p-value is obtained by referring F to an F distribution on (n 1, n 2 ) df. For nonnormal models, the p-value is obtained by referring D k D m to a χ 2 distribution on 2m 2k df. These p-values for comparing models are approximate and are typically somewhat conservative (Royston and Altman 1994).
24 fp Fractional polynomial regression Acknowledgment We thank Patrick Royston of the MRC Clinical Trials Unit, London, and coauthor of the Stata Press book Flexible Parametric Survival Analysis Using Stata: Beyond the Cox Model for writing fracpoly and fracgen, the commands on which fp and fp generate are based. We also thank Professor Royston for his advice on and review of the new fp commands. References Becketti, S. 1995. sg26.2: Calculating and graphing fractional polynomials. Stata Technical Bulletin 24: 14 16. Reprinted in Stata Technical Bulletin Reprints, vol. 4, pp. 129 132. College Station, TX: Stata Press. Isaacs, D., D. G. Altman, C. E. Tidmarsh, H. B. Valman, and A. D. Webster. 1983. Serum immunoglobulin concentrations in preschool children measured by laser nephelometry: Reference ranges for IgG, IgA, IgM. Journal of Clinical Pathology 36: 1193 1196. Libois, F., and V. Verardi. 2013. Semiparametric fixed-effects estimator. Stata Journal 13: 329 336. Royston, P. 1995. sg26.3: Fractional polynomial utilities. Stata Technical Bulletin 25: 9 13. Reprinted in Stata Technical Bulletin Reprints, vol. 5, pp. 82 87. College Station, TX: Stata Press. Royston, P., and D. G. Altman. 1994. Regression using fractional polynomials of continuous covariates: Parsimonious parametric modelling. Applied Statistics 43: 429 467. Royston, P., and G. Ambler. 1999a. sg112: Nonlinear regression models involving power or exponential functions of covariates. Stata Technical Bulletin 49: 25 30. Reprinted in Stata Technical Bulletin Reprints, vol. 9, pp. 173 179. College Station, TX: Stata Press.. 1999b. sg81.1: Multivariable fractional polynomials: Update. Stata Technical Bulletin 49: 17 23. Reprinted in Stata Technical Bulletin Reprints, vol. 9, pp. 161 168. College Station, TX: Stata Press.. 1999c. sg112.1: Nonlinear regression models involving power or exponential functions of covariates: Update. Stata Technical Bulletin 50: 26. Reprinted in Stata Technical Bulletin Reprints, vol. 9, p. 180. College Station, TX: Stata Press.. 1999d. sg81.2: Multivariable fractional polynomials: Update. Stata Technical Bulletin 50: 25. Reprinted in Stata Technical Bulletin Reprints, vol. 9, p. 168. College Station, TX: Stata Press. Royston, P., and W. Sauerbrei. 2008. Multivariable Model-building: A Pragmatic Approach to Regression Analysis Based on Fractional Polynomials for Modelling Continuous Variables. Chichester, UK: Wiley. Smith, J. M., C. J. Dore, A. Charlett, and J. D. Lewis. 1992. A randomized trial of Biofilm dressing for venous leg ulcers. Phlebology 7: 108 113. Also see [R] fp postestimation Postestimation tools for fp [R] mfp Multivariable fractional polynomial models [U] 20 Estimation and postestimation commands